Earth and Planetary Science Letters 253 (2007) 429 – 438

www.elsevier.com/locate/epsl

Earthquake scaling, fault segmentation, and structural maturity

Isabelle Manighetti ⁎, Michel Campillo, Sylvain Bouley, Fabrice Cotton
Laboratoire de Géophysique Interne et Tectonophysique de Grenoble, Grenoble, France
Received 28 July 2006; received in revised form 17 October 2006; accepted 2 November 2006
Available online 6 December 2006
Editor: R.D. van der Hilst

Abstract

Slip and length measurements on earthquakes suggest large stress drop variability. We analyze an extended set of slip-length
measurements for large earthquakes (M ≥ 6) to seek for the possible origin(s) of this apparent variability. We propose that such
variability arises from earthquakes breaking a variable number of major fault segments. That number depends on the strength of the
inter-segment zones, which itself depends on the structural maturity of the faults. We propose new Dmax–L parameterizations based
on that idea of multiple segment-ruptures. In such parameterizations, each broken segment roughly scales as a crack, while the total
multi-segment rupture does not. Stress drop on individual segments is roughly constant, only varying between 3.5 to 9 MPa. The
slight variation that is still observed depends on fault structural maturity; more mature faults have lower stress drops than immature
ones. The new Dmax–L functions that we propose reduce uncertainties with respect to available relationships. They thus provide a
more solid basis to estimate seismic hazard by integrating fault properties revealed by geological studies.
© 2006 Elsevier B.V. All rights reserved.

Keywords: Earthquakes osin; Scaling; Slip-length; Fault maturity; Fault segmentation

1. Introduction behavior was real (see stormy debate that opposed

A common hypothesis in earthquake mechanics is
that earthquakes have a macroscopic behavior of cracks
in an elastic medium, with the stress drop being a
material property hence being almost constant for
crustal earthquakes e.g., [1–5]. From that hypothesis,
one expects maximum (Dmax) or mean displacement
(Dmean) on earthquakes to scale with rupture length (L)
when L ≤ 2Wseism (Wseism being the thickness of the
seismogenic layer), and tapers off for long ruptures
(L N 2Wseism). While available D–L earthquake data
have long been too few to show whether or not that
⁎ Corresponding author.
E-mail address: imanighe@obs.ujf-grenoble.fr (I. Manighetti).
0012-821X/$ - see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.epsl.2006.11.004
‘L- and W-model’ supporters over 15 yr; [6–12]), it is
now generally admitted (exceptions are [13,14]), from
examination of denser data sets, that earthquakes more
likely follow a ‘W-model’, hence roughly behave as
cracks [15–17]. Shaw and Scholz (2001; later referred to
as S&S01) have recently proposed a scale-invariant
physical model that includes the two D–L regimes:

for ruptures with LV2W ; Dmean ¼ a⁎ðL=2Þ

for ruptures with LN2W ; Dmean ¼ a⁎ð1=½1=L ð1Þ
þ1=2W Þ
It is important to note that, in the formulation of that
equation, S&S01 postulate that α is proportional to a
constant static stress drop, while W represents the width
430 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438

of the seismogenic layer thus is Wseism. The model appropriate overall, a closer examination of the data
roughly reproduces the first-order distribution pattern of suggests that an additional factor may intervene in the
∼ 55 earthquake slip-length data (see their Fig. 1). While D–L scaling for actual D–L data are much more
this suggests that the physical basis of the model is scattered than predicted by the constant stress drop

Fig. 1. Displacement-length surface data in Asia (from Table 2). (a) Dmax versus L plot. Symbol size is proportional to quality weight. (b) Fit quality
map resulting from adjusting the data set with one S&S01' function having variable [α,W ] values. Three, and possibly fours regions of [α,W ] values
are found that best-fit distinct data subsets. (c) Data are plotted together with the three optimal S&S01' functions deduced from B. Corresponding
optimal [α,W ] values are indicated.
 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438
model; large earthquakes of similar size (i.e., similar
length and width) obviously can produce different slips
and stress drops, as it has been pointed out by many
authors e.g., [2,17–21]. We focus here on such slip
variability and seek for its possible origin(s). We
compile all available slip-length measurements for
large earthquakes in four seismically active regions of
the world. Our data set only includes earthquakes with a
magnitude larger than ∼ 6, so that our results are
relevant to the largest scale of the earthquake process
only. We conduct our analysis using the functional form
of Eq. (1). Yet, we give an interpretation of the para-
meters α and W that is different, though not contradic-
tory, with that of S&S01. We state that the observed slip
variability arises from broken geological faults having a
variable frictional strength that depends on their long-
term slip history (‘structural maturity’) and geometry
(large-scale segmentation). While the effect of slip
history has already been evoked in earthquake scaling
analyses e.g., [18,22–25], that of fault segmentation has
never been. We claim that it should be, for it dictates the
way earthquakes gain in length, hence strongly governs
the relationship(s) between earthquake slip and length.
2. Data sets
We examine displacement-length data for a set of
∼ 250 large (M ≥ ∼ 6), shallow (rupture width ≤ 40 km,
with an average value Wmean of 18 km), continental
earthquakes of mixed focal mechanisms (strike–slip,
reverse and normal), that have occurred in four of the
most seismically active regions worldwide: Asia (broad
sense), Turkey, West US, Japan. In these regions, long-
term active faults are generally well known, with their
surface geometry (total length, segmentation, strike
variations, associated secondary fault networks), age,
maximum slip rate and total cumulative displacement
being generally determined. We use these long-term
parameters (where available) to qualify the structural
maturity of the faults that broke during the analyzed
earthquakes, as explained in Table 1 (supporting online
material). Doing so, we classify the broken faults in
three classes, basically ‘immature’, ‘intermediate’, and
‘mature’ (Table 1). Earthquake slip-length data are
compiled from literature (references in Tables 2 and 3;
supporting online material). We consider here the
maximum displacement (Dmax), not the mean (Dmean),
for it is best constrained. Besides, Manighetti et al. [21]
have shown that Dmax = 2 ⁎ Dmean for most large earth-
quakes worldwide, a property that is found to be scale-
independent. Being aware that rupture slip and length
measured at surface may be lower than actual maximum
431
slip and length produced at depth, we compile both
surface measurements (209 earthquakes; Table 2) and
slip-length values inferred at depth from earthquake
source models (56 earthquakes; Table 3), and analyze
them separately. Surface data are screened for quality,
and a ‘quality weight’ assigned to each data from criteria
defined in Table 2. Note that, since error bars cannot be
properly defined for most data, weighting them with a
quality factor is the best we can do to discriminate poor
and robust data. While such quality screening is a
fundamental step to discuss any scaling law, it has never
been done before (in terms of quality weighting of each
data). Concerning slip-length data at depth, we did not
attempt to ‘qualify’ the quality of the various earthquake
source models. Rather we chose to average the different
Dmax–L values proposed for the same earthquakes
(Table 3).
3. Data analysis
We start analyzing in detail the Asian data set, for it is
the densest. Fig. 1a shows the Asian surface Dmax–L
data, with symbol size proportional to quality weight.
The data are rather dispersed, so that a single function
cannot be found to adjust them all properly. However
the overall shape of the data distribution resembles an
asymptotic function similar to that predicted by the
S&S01' model; slip increases with length for short
ruptures (L b 100–200 km), and seems to saturate for
longer ruptures. We note that, while data are dispersed,
they do not extend evenly over the plot; there are zones
free of data, and zones where data cluster. This suggests
the existence of a few specific trends. To check whether
or not these trends are significant, we use a figure of
merit M to score the adjustment of a couple of [α, W ]
parameters to the data set, and explore the whole space
of the model parameters (Fig. 1b). M is of the form
M = ∑i(ch([Dobs − D(α,W )] / Dn))− 1 where Dobs is the
measured value of slip, D(α,W ) is the predicted value
of slip, and Dn is an adjustable smoothing parameter
(chosen here to be 7 times the standard deviation of
Dobs). At this stage, we thus assume that α and W are
free adjustment parameters. The existence of several,
distinct zones of maximum in Fig. 1b indicates that the
data cannot be fitted with a single model, but rather
include several subsets associated with different models,
i.e. different values of [α, W ]. Three major zones are
identified, whose shape is related to bias between α and
W. Since we deal with a limited number of data, we must
verify that this multimodal structure is not fictitious. We
generated random models having the statistical proper-
ties of the data (average and variance of D depending on
432 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438

L, distribution of L) and performed the same test as dozen realizations to reach this conclusion. Examples
before. The multimodal structure with respect to the are presented in Fig. A (supporting online material). We
functional form of S&S01 that is exhibited by the data is can thus conclude on firm ground that the Asian surface
not produced by sparse random series. We checked a data set actually includes at least 3 distinct groups

Fig. 2. Displacement-length surface data in the four Asia, West US, Turkey and Japan regions (from Table 2). (a) Dmax versus L plot. Symbol size is
proportional to quality weight. (b) Fit quality map resulting from adjusting the whole data set (smallest ruptures with L b 50 km are excluded for
clarity) with one S&S01' function having variable [α,W ] values. Three, and possibly fours regions of [α,W ] values are found that best-fit distinct data
subsets. (c) Data are plotted together with the four optimal S&S01' functions deduced from b. Corresponding optimal [α,W ] values are indicated.
 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438 433

unambiguously associated with different couples of [α,
W ] parameters. Based on these findings, we searched
for the three optimal S&S01' functions that combine to
produce the best fit to the entire data set. This is done
through an iterative optimization procedure in which
each data point is affected to a subset. Fig. 1c shows the
results, together with the three couples of optimal [α, W ]
values best adjusting the data. We find that data are
properly fitted provided that W and α both vary. In the
last section of the paper, we discuss statistical tests
(using Akaike Information Criterion) that show that the
misfit reduction is not due to ‘over-fitting’ the data with
models having a too large number of free parameters.
The optimal values of W are found to be regularly
spaced, the largest being about multiple of the lowest,
while α varies irregularly yet decreasing with W. Note
that these results are independent of slip mode; strike–
slip, reverse and normal ruptures are found in any of the
three groups (Fig. B, done for all data; supporting online
material).
We applied the same treatment to the other three
regional data sets. In all cases the data show a multimodal
structure, with at least two or three distinct regions of [α,
W ] values found to adjust distinct data subsets (Fig. C,
supporting online material). Fig. D (supporting online
material) shows the corresponding optimal S&S01' func-
tions. In all cases, the optimal [α, W ] parameters vary in a
discrete, almost regular fashion. Besides, the optimal [α,
W ] values are roughly similar from one region to the
other, so that, on average, 4 recurrent, multiple couple
values are suggested, equal to [(46 ± 6).10− 5, 17.5 ± 0.5],
[(20± 2).10− 5, 32 ± 3], [(8 ± 1).10− 5, 57 ± 4], and [(5 ±
1).10− 5, 71.5 ± 1.5].
Fig. 2a now shows all surface data together. The
uneven data distribution appears more clearly, with three
main trends distinguishable. Fig. 2b confirms that the
whole data set consists of three or four distinguishable
data subsets associated with distinct domains of [α, W ].
Fig. 2c shows the four optimal S&S01' functions that we
inferred from this basis. These functions again suggest

Fig. 3. Proposed scenario. Earthquakes break a variable number of major segments along the faults on which they initiate. That number increases with
the degree of structural maturity of the faults, for inter-segments zones have lower strength on mature faults. The shape and amplitude of the resulting
slip profiles varies accordingly.
434 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438
discrete values of W, roughly equivalent to the lowest
18 km value being multiplied by 1 to 4 when one goes
from optimal functions 1 to 4. The parameter α is also
found to vary in a discrete, yet more irregular fashion.
Finally, Fig. F (supporting online material) shows all
slip-length data at depth deduced from inversion
models, together with the four optimal S&S01' func-
tions whose existence is inferred, as before, from
analysis of Fig. E (supporting online material). Though
the data set is different, the optimal W values are similar
to those obtained from surface data, while the optimal α
values are slightly higher.
Together these results highlight major issues. First, the
finding that all data sets contain several distinct data
subsets, hence request the combination of several distinct
Dmax–L functions to be adjusted overall, demonstrates
that an additional parameter, other than Wseism and a
constant stress drop, intervenes in the relationship
between rupture slip and length (note that this makes
that relationship not being strictly speaking a ‘scaling
law’). Second, the optimal S&S01' functions that are
inferred imply variable values of W. Thus, the single crack
model does not apply in its simple form. Besides, the
range of variation of W is large, while the obtained values
are roughly multiple of the lowest. Together these are
incompatible with W being the seismogenic thickness.
We thus need to admit that W does not represent the width
of the seismogenic zone. Rather it is a characteristic length
at which slip saturates, to which we will give an inter-
pretation in the following. Finally, the optimal functions
that we find also imply variable values of α, thus of stress
drop, given by Δσ = μ ⁎ (α / 2) with μ = 3.1010 in the
original S&S01 interpretation. For instance, when the
stress drop is calculated from the inversion model data
(total slip values, Fig. F), these variations appear to be
quite large, ranging between ∼9 and 1 MPa as one goes
from optimal functions 1 to 4. Such large variations are
incompatible with the constant stress drop hypothesis.
This further confirms that the crack model scaling does
not apply to earthquakes in its simple form.
4. Interpretation
This led us to seek for a scenario that would retain
basic, reasonable ingredients such as the elastic crack
behavior and the reality of the seismogenic thickness
(Wseism ∼18 km for our data), while allowing the
parameter W in Eq. (1) to take discrete values increasing
algebraically. As said before, that fitting parameter W is
not Wseism, and is unlikely to represent the thickness of
anything real for the obtained optimal values do not
coincide with any known feature of the Earth structure.
We thus hypothesize that W rather is the half rupture
length at which slip starts saturating (Lsat); we actually
observe that slip starts saturating at different lengths along
the analyzed ruptures. Some earthquakes would behave as
‘simple cracks’, having Lsat = 2Wseism, while some earth-
quakes would be more complex, with their slip starting to
saturate at greater lengths observed to be 2, 3 or 4 times
longer than that for a single crack (Lsat = n ⁎ (2Wseism) with
n between 2 and 4). This behavior suggests that those
earthquakes are made of several cracks juxtaposed along
the rupture strike. This recalls fault segmentation. Indeed,
geological faults are segmented, made of a finite number
of large-scale segments [26–29]. Note that we are
referring here to the segmentation that affects a fault at
its largest scale; not to ‘slip heterogeneity’ on fault planes
as commonly described in seismological literature e.g.,
[30]. The first-order segments that we are evoking thus are
of about the same scale than the whole fault they belong to
(at most shorter by one order). Generally, only 3–4 such
large-scale segments are identified along a fault, inde-
pendent of its slip mode e.g., [26,27,29,31,32]. As a fault
grows with time, linkage between its large-scale segments
evolves from ‘soft’ (i.e., segments are hardly linked) to
‘hard’ (i.e., segments are fully linked) e.g., [33–36], so
that the geometry of the fault zone simplifies and becomes
more continuous, more ‘through-going’ [26,28,37]; inter-
segment zones evolve from being wide areas of
distributed, disorganized, secondary fissuring and fault-
ing, to becoming narrow zones of localized, through-
going faulting. We thus expect the zones that connect the
large-scale segments to have an apparent strength that
depends on the structural maturity of the overall fault they
belong to; high on young, immature faults, and lower on
long-lived, mature features [28,29]. Large-scale inter-
segment areas may thus behave as more or less ‘resistant
barriers’ to earthquake rupture ‘propagation’, as has
actually been observed in many cases e.g., [21,38–42].
We propose that, depending on the strength of these inter-
segment barriers, an earthquake may eventually break a
variable number of large-scale segments along the fault on
which it initiates. On mature faults, the breakage of a first
segment may easily overcome the resistance of the inter-
segment barriers, so that it may trigger the cascading
rupture of several segments along the fault, resulting in a
long, multi-segment rupture. By contrast, on immature
faults, the breakage of a first segment is unlikely to
overcome the barriers that disorganized, wide inter-
segment zones represent, so that only one segment may
eventually break, resulting in a short, crack-like rupture.
Fig. 3 shows the scenario that we envision. In that
scenario, segments have the same length and width (equal
to Wseism for large earthquakes), while the size of barriers
 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438 435

is neglected. Segments behave independently, so that each
breaks as an elastic crack. Thus, if only one fault segment
breaks, slip starts saturating at Lsat = 2Wseism. But if the
inter-segment barriers are ‘weak enough’ to allow
cascading triggering, two, three or more segments along
the fault may break in succession, leading to increase Lsat
by as many times as there are broken segments (Fig. 3).
5. Facing the model with independent observations
We now test that hypothesis further. A first test
consists in facing the actual number of major fault
segments broken during earthquakes to the four optimal
functions determined earlier (Fig. 4a). We found such
information for ∼40 earthquakes (details in Table 2).
Fig. 4a shows that, overall, earthquakes that have
broken one or two major segments pertain to the data
subsets that are best fitted by the optimal functions 1 or
2, while earthquakes that have broken 3 to 4 segments
pertain to the data subsets that are best fitted by the
optimal functions 3 or 4. This is in keeping with the
scenario that we propose (Fig. 3).
Another piece of information comes from Manighetti
et al. [21]. These authors have studied the generic
properties of earthquake slip profiles and shown that
these profiles have a self-similar triangular shape that is
roughly symmetric when a single major fault segment
has been broken, and asymmetric to a various degree

Fig. 4. Facing the segmentation scenario with independent data. (a) Number of broken segments as a function of proximity with optimal functions 1 to
4; the number of broken segments is observed to increase as one goes from function 1 to 4. (b) Asymmetry of surface slip profiles (from Table 2; inset
recalls the 3 classes of asymmetry depicted by Manighetti et al., 2005) as a function of proximity with optimal functions 1 to 4; asymmetry is observed
to increase as one goes from function 1 to 4. (c) Asymmetry of along-strike slip profiles at depth (from Table 3; inset as in b) as a function of proximity
with optimal functions 1 to 4; asymmetry is observed to increase as one goes from function 1 to 4. (d) Structural maturity of broken geological faults
(from Table 1) as a function of proximity with optimal functions 1 to 4; fault maturity is observed to increase as one goes from function 1 to 4.
436 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438

Fig. 5. Modeling the slip-length data. (a) Surface Dmax–L data (from Table 2). (b) Dmax–L data inferred at depth (from Table 3). Each global data set is
shown with the theoretical functions deduced from the multi-segment rupture model. The regression of Wells and Coppersmith [19] is shown in A for
comparison. Apparent stress drops vary from ∼ 9 to 1 MPa as one goes from function 1 to 4 (in b). Yet, actual stress drop on individual crack-like
segments only vary from ∼ 9 to 3.5 MPa.

when several segments have been broken; in that later
case, the segment with higher slip occupies approxi-
mately a third of the total rupture length. Fig. 4b–c
report the degrees of asymmetry of ∼90 earthquake slip
profiles as defined by [21], measured at surface (Fig. 4b)
or inferred at depth from inversion models (Fig. 4c). The
data are presented with respect to the four optimal
functions defined before. Overall, symmetric slip
profiles are found for earthquakes pertaining to the
data subsets best fitted by the optimal functions 1 or 2,
while more asymmetric profiles are found for earth-
quakes pertaining to the data subsets best fitted by the
optimal functions 3 to 4.
We have proposed that the strength of the large-scale
inter-segment barriers depends on the structural maturity
of the faults to which segments belong, with such
barriers being ‘stronger’ on immature faults, and
‘weaker’ on mature faults. Fig. 4d faces the structural
maturity (as defined in Table 1) of the faults broken in
140 earthquakes, with the four optimal functions
established from surface data. It confirms that overall,
earthquakes occurring on immature faults pertain to
groups 1 or 2, whereas earthquakes occurring on most
mature faults pertain to groups 3 or 4.
6. Discussion and conclusions
Different pieces of evidence thus converge to suggest
that the scenario proposed in Fig. 3 represents a valuable
basis to interpret the earthquake D–L relationships. An
earthquake would break a variable, yet limited number
of major segments (or major ‘asperities’) along a fault,
 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438
depending on whether or not the breakage of the very
first segment is ‘energetic’ enough to overcome the
strength of the first encountered major inter-segment
barrier, and trigger the rupture of the neighboring major
segment(s). That capability of breaking inter-segment
barriers would depend on the fault structural maturity;
mature faults would be more prone to break in cascading
events for their inter-segment barriers are quite narrow
and smoothed, whereas more immature faults would
more likely break in a single or double event for their
inter-segment zones represent much stronger barriers to
rupture propagation. The variability in displacement-
length earthquake data would result from that multiple
event rupturing process. Earthquakes breaking a single
segment would behave as a single elastic crack with slip
starting to saturate at L = 2Wseism, while earthquakes
breaking several segments would behave as a juxtaposition
of several cracks. As a consequence, the minimum rupture
size at which slip starts saturating is different from one
earthquake to the other, yet in the form Lsat =n ⁎ (2Wseism)
with n the number of broken segments when those are
assumed to be similar in length. Intermediate situations
with segments of different lengths are likely to exist and
result in the scatter of the data points in each subset (in
addition to stress drop variability). Yet, it is noteworthy that
a model with Lsat being multiple of the seismogenic
thickness is actually most required by the data. Fig. 5 faces
the available Dmax–L data (surface: Fig. 5a; depth: Fig. 5b)
to that theoretical model (defined for Wseism = 18 km). The
adjustment is satisfying: the standard deviation of the
residuals log(Dobs)–log(Dpredicted) calculated as in [19] is
only 0.18. The stress drop variations that are suggested by
the α variations are not as large as it seems when
neglecting segmentation. Indeed, for a data subset
corresponding to a n-segment rupture, the stress drop on
each segment is Δσ =n ⁎ μ ⁎ (α/2). The inferred stress
drops are actually about the same on all broken segments,
on the order of 3.5–9 MPa (calculated from inversion
model data which are more characteristic of the ruptures at
depth; Fig. 5b). The few MPa difference that is still
observed obviously depends on fault structural maturity
(Fig. 4d); segments on mature, hence weakened faults
break in lower stress drop-earthquakes than segments on
immature faults, as has been suggested before e.g.,
[22,29,37]. Note that we verified that this variation in
apparent stress drop is not related with the effect of the
finite size of the barriers between ruptured segments.
Whatever that size (in a realistic range), a change in stress
drop is required. The data suggest stress drops ranging
from ∼9 MPa for single segment-events on immature
faults, to ∼3.5 MPa for 4 segment-events on mature faults
(Fig. 5b). We thus broaden the conclusions of Cao and Aki
437
(1986) and Anderson et al. (1996), stating that fault
structural maturity is a major factor, if not the most
important, that governs the stress drop of earthquakes,
hence the amplitude of ground motions. Careful geological
analyses of long-term faults should thus be included in any
seismic hazard study. Fig. 5 shows that the maximum
length of earthquakes is ∼200 km for group 1, ∼400 km
for group 2, ∼500 km for group 3, suggesting that major
segments along worldwide faults are ∼200 km-long.
Together with the observation that stress-drop and
segmentation are related, this explains why Dmax tends to
decrease with total length for large events as pointed out by
[30].
Our reappraisal of displacement-length earthquake
data has important implications on seismic hazard
assessment. First, depending on their structural maturity,
faults of similar length may produce significantly
different amounts of slip; mature faults obviously
break in long ruptures with low slip amplitudes (b 4–
7 m). By contrast, more immature faults break in shorter,
yet more ‘energetic’ ruptures, on which slip as high as
15 m may be expected.
On the other hand, when one calculates the ratio
between maximum slip measured at surface and
maximum slip inferred at depth, that one finds that
surface slip generally is only a fraction of actual slip at
depth, averaging 85% for most large strike–slip ruptures
(M ≥ 6.5), 40% for small (∼6 ≤ M b 6.5) strike–slip,
reverse and normal earthquakes, and even less than 10%
for a few other cases (Fig. G, supporting online
material). This suggests that surface measurements of
slip cannot be used to calculate an earthquake
magnitude unless they are ‘corrected’ by a certain factor.
Finally, our refined D–L earthquake ‘scaling laws’
are more accurate than those available (see in Fig. 5a a
comparison with regression from Wells and Copper-
smith, 1994; later referred to as W&C). We have
calculated the standard deviation of the residuals log
(Dobs)–log(Dpredicted) for our data set, using the W&C
empirical relation log(Dpredicted) = − 1.38 + 1.02 log(L).
The standard deviation that we obtain is equal to 0.41, a
value similar to the one obtained in the original paper of
W&C from a smaller data set. Table 4 (supporting
online material) indicates however that the residuals
obtained with the W&C relation show a clear bias with
earthquake size; the W&C model overestimates slip
amplitudes for earthquakes with lengths greater than
200 km (see also Fig. 5a). To further compare our
regressions with that of W&C, we have also calculated
the residuals that would arise from fitting our entire data
set with one single optimal S&S01 relation (that is
found to have W = 54.7 km and α = 12.4 10− 5). The
438 I. Manighetti et al. / Earth and Planetary Science Letters 253 (2007) 429–438
variance is similar to the one obtained when using the
W&C relation (σ = 0.41), but the residuals do not exhibit
any bias with earthquake size. The residuals also ap-
proximately follow a log-normal distribution (Fig. H,
supporting online material).
The standard deviation of the residuals obtained when
using our best-fitting model with four curves is 0.18
when the entire data set is considered. For the best
constrained Asian data set whose modeling requires only
three curves, the standard deviation keeps as low (0.17).
We checked that this decrease in standard deviation is not
an ‘artefact’ due to the introduction of a too large number
of parameters (over-fitting). This was done by comput-
ing the Akaike Information Criterion AIC; [43]. This
criterion determines the balance between the model
improvement and the number of free parameters that
contribute to that improvement (increasing the number
of free parameters always improves the fit). We com-
puted the AIC for models with n curves and verified that
that criterion is decreasing continuously with n, for n
between 1 and 4. This test shows that our 4 curve-model
significantly improves the residuals without over-fitting.
This confirms the potential interest of modelling earth-
quake slip-length data with a multiple event-model of the
form we propose.
In hazard studies where a critical issue is to reduce
the uncertainties on seismic hazard assessment e.g. [44],
our study may thus have some impact since it produces
residues that are twice lower than those associated with
available D–L regressions [19]. Our predictive Dmax–L
functions could be used as distinct branches of a logic
tree e.g., [45], each with an reduced aleatory variability
compared to classical models. Weight assigned to each
branch would depend on fault structural maturity
(epistemic uncertainty). We are now in the process of
refining the ‘maturity criteria’, so that more accurate
functions may be established between earthquake slip
and length, and fault structural maturity.
Acknowledgements
We thank M. Cocco, R. van der Hilst and an
anonymous reviewer for their comments that helped
improving the paper. We thank Y. Delaby, B. Valette and
N.A. Abrahamson for their useful suggestions for the
data processing.
Appendix A. Supplementary data
Supplementary data associated with this article can
be found, in the online version, at doi:10.1016/j.epsl.
2006.11.004.
References
[1] K. Aki, Tectonophysics 13 (1972) 423.
[2] H. Kanamori, D.L. Anderson, Bull. Seismol. Soc. Am. 65 (1975)
1073.
[3] J.N. Brune, Seismic Risk and Engineering Decisions, 1976, p. 140.
[4] R. Madariaga, Bull. Seismol. Soc. Am. 66 (1976) 636.
[5] H. Kanamori, E.E. Brodsky, Rep. Prog. Phys. 67 (2004) 1429.
[6] C. Scholz, Bull. Seismol. Soc. Am. 72 (1982) 1.
[7] B. Romanowicz, Geophys. Res. Lett. 19 (1992) 481.
[8] B. Romanowicz, J. Rundle, Bull. Seismol. Soc. Am. 83 (1993) 1294.
[9] C. Scholz, Bull. Seismol. Soc. Am. 84 (1994) 215.
[10] C. Scholz, Bull. Seismol. Soc. Am. 84 (1994) 1677.
[11] B. Romanowicz, Bull. Seismol. Soc. Am. 84 (1994) 1675.
[12] P. Bodin, J.N. Brune, Bull. Seismol. Soc. Am. 86 (1996) 1292.
[13] J.H. Wang, S.S. Ou, Bull. Seismol. Soc. Am. 88 (1998) 758.
[14] T.C. Hanks, W.H. Bakun, Bull. Seismol. Soc. Am. 92 (2002) 1841.
[15] P.M. Mai, G.C. Beroza, Bull. Seismol. Soc. Am. 90 (2000) 604.
[16] B.E. Shaw, C.H. Scholz, Geophys. Res. Lett. 28 (2001) 2991.
[17] B. Romanowicz, L.J. Ruff, Geophys. Res. Lett. 29 (2002) 12.
[18] C. Scholz, C.A. Aviles, S.G. Wesnousky, Bull. Seismol. Soc.
Am. 76 (1986) 65.
[19] D.L. Wells, K.J. Coppersmith, Bull. Seismol. Soc. Am. 84 (1994)
974.
[20] D.J. Dowrick, D.A. Rhoades, Bull. Seismol. Soc. Am. 94 (2004)
776.
[21] I. Manighetti, M. Campillo, C. Sammis, M. Mai, G. King, J. Geo-
phys. Res. 110 (2005) B05302.
[22] T. Cao, K. Aki, Pure Appl. Geophys. 124 (1986) 515.
[23] J.G. Anderson, S.G. Wesnousky, M.W. Stirling, Bull. Seismol.
Soc. Am. 86 (1996) 683.
[24] C. Marone, Annu. Rev. Earth Planet. Sci. 26 (1998) 643.
[25] S.A. Miller, Geophys. Res. Lett. 29 (2002), doi:10.1029/
2001GL014181.
[26] M. Stirling, S.G. Wesnousky, K. Shimazaki, Geophys. J. Int. 124
(1996) 833.
[27] I. Manighetti, G.C.P. King, Y. Gaudemer, C. Scholz, C. Doubre,
J. Geophys. Res. 106 (2001) 13667.
[28] Y. Ben-Zion, C.G. Sammis, Pure Appl. Geophys. 160 (2003) 677.
[29] K. Otsuki, T. Dilov, J. Geophys. Res. 110 (2005), doi:10.1029/
2004JB003359.
[30] J. Liu-Zeng, T. Heaton, C. DiCaprio, Geophys. J. Int. 162 (2005) 841.
[31] D.P. Schwartz, K.J. Coppersmith, J. Geophys. Res. 89 (1984) 5681.
[32] L.J. An, C.G. Sammis, Tectonophysics 253 (1996) 247.
[33] B. Trudgill, J. Cartwright, GSA Bull. 106 (1994) 1143.
[34] J.J. Walsh, J. Watterson, W.R. Bailey, C. Childs, J. Struct. Geol.
21 (1999) 1019.
[35] M.J. Young, R.L. Gawthorpe, S. Hardy, J. Struct. Geol. 23 (2003)
1933.
[36] N.J. Richardson, J.R. Underhill, G. Lewis, Basin Res. 17 (2005) 203.
[37] S.G. Wesnousky, Nature 335 (1988) 340.
[38] K. Aki, J. Geophys. Res. 84 (1979) 6140.
[39] G. King, J. Nabelek, Science 228 (1985) 984.
[40] R.H. Sibson, Nature 316 (1985) 248.
[41] R.A. Harris, S.M. Day, J. Geophys. Res. 98 (1993) 4461.
[42] R.A. Harris, S.M. Day, Geophys. Res. Lett. 26 (1999) 2089.
[43] H. Akaike, IEEE Trans. Automat. Control 19 (1974) 716–723.
[44] F. Scherbaum, J. Bommer, H. Bungum, F. Cotton, N.A. Abrahamson,
Bull. Seismol. Soc. Am. 95 (2005) 1575.
[45] J. Bommer, F. Scherbaum, H. Bungum, F. Cotton, F. Sabetta, N.A.
Abrahamson, Bull. Seismol. Soc Am. 95 (2005) 377.